HPC Cloud TCOs

In general, on premises HPC resources must be over-provisioned in order to meet peak needs. It is intuitively obvious that it should be possible to provision on premises HPC resources in a leaner manner, and then use public clouds to meet peak demands. Unfortunately, from an HPC pricing perspective, cloud resources are quite expensive – frequently 10x more per hour than large scale, cost optimized, on premises compute. This makes setting that lean provisioning policy in an optimal, low risk manner quite difficult.

This presentation provides an overview for a mathematical framework for HPC hybrid cloud deployments:

• Provides a quantitative total cost function for HPC hybrid cloud deployments
• The total cost function may be optimized using standard optimization algorithms to yield the lowest cost solution
• Being based on historical and live cluster data, and the cost function may be automatically tuned as HPC loading conditions change over time
• The cost function naturally comprehends loading variability, and provides not just a cost number but a 90% confidence interval
• Not only are robust confidence intervals provided, but an actuarial risk assessment may also be derived from the computation – we we can quantify the risk of overspend